SecRL-Prune: RL-Based Pruning of Code LLMs that Preserves Adversarial Code Mutation

arXiv·medium signal

Code LLMs can perform functionality-preserving code mutation that adversaries use to spin up malware variants evading signature detection. SecRL-Prune studies structured RL-based pruning of code models and the security implication that compression can preserve — or even concentrate — this mutation capability. A dual-use reminder for teams shipping or distilling code models. Published 2026-06-04 (cs.CR/cs.SE).